The vocabulary report on the owner's store (2026-10-02)

A dated snapshot, as of 2026-10-02. An ontology is built from evidence: which relations occur, how often, what values they take, which kinds of thing they attach to, which entities line up with an external knowledge base, and which modelling choices a person has confirmed. A store of beliefs holds all of that already. This page runs the report that reads it out, particles query --vocabulary, on a read-only copy of the owner's live store, and checks that it reproduces the figures first measured by hand with SQL on 2026-09-30.

What the report computes

The report is computed on every call and never stored. It has two parts.

A store header.

• Subjects in total, the subjects aligned to each external namespace, and, within a namespace, how many fall in each link-confidence band. The bands are the ones the subject resolver already records: asserted (confidence 1.0, from a structured extractor or an operator), scored (an embedding score at or above subjects.wikidata_link_suppress_threshold), unscored (the 0.5 a link gets when no content could score it), and suppressed (below the threshold, stored but hidden from export and query output).
• Subjects carrying a class, by class and by the namespace the class is minted in.
• Structured claims in the store, the claims in view after the query's default exclusions, and their object kinds and value shapes.
• The confirmed modelling decisions in the operator event log, by event type: merged or unmerged duplicates, closed contradictions, resolved reviews, and subject merges, splits, aliases, reclassifications and link rulings.

One row per canonical predicate. Surface forms are grouped under the deterministic canonical form that predicate profiles compare, so has status, had status and have status are one row with all three listed. Each row carries its claim count, its object value shapes (text, numeric, dated, URI, token), the classes of the subjects its claims are about, and its alignment: the term an adopted vocabulary document mints or aliases for the predicate, and that term's links to external properties. The alignment column is empty on every row of this run, since the store has adopted no vocabulary document, and the report says so rather than leaving the column out.

A value shape is a reading of form, for inferring a datatype: an untyped literal "1987" counts as numeric. The structural query filters still compare only typed values.

The report reads claims through a project observer when one is named, like every other read that selects beliefs. The subject header and the decision counts are store-wide, since subjects and operator events are not beliefs a project observes. With --as-of, the claims are those believed at the instant, the header counts the subjects created by then (with their classes and links as they stand now), and the decisions are those recorded by then.

How it was run

On a copy of the store taken with SQLite's backup interface from a read-only connection:

DATABASE_URL=sqlite+aiosqlite:///store-copy.db uv run particles query --vocabulary
DATABASE_URL=sqlite+aiosqlite:///store-copy.db uv run particles query --vocabulary --as-of 2026-10-01
DATABASE_URL=sqlite+aiosqlite:///store-copy.db uv run particles query --vocabulary \
    --as-of 2026-10-01 --include-document-meta --include-non-asserted

Each run took about eight seconds, with no embedding and no LLM call. The same report is served as JSON by --format json and by GET /vocabulary.

The figures first measured, reproduced

The 2026-09-30 measurement was taken by hand before midnight UTC. Reading the store as of 2026-10-01 at 00:00 UTC reproduces every figure.

Figure	Measured 2026-09-30	Report as of 2026-10-01
Distinct predicates	6,358	6,358
Structured claims the listing counts	11,074	11,074
ACTIVE structured claims, no default exclusions	11,462	11,462
Literal objects among them	11,338	11,338
URI objects among them	124	124
Subjects	9,161	9,161
Subjects aligned to Wikidata	582	582
Subjects classed in the local artifact: namespace	2,243	2,243
DUPLICATES_MERGED	181	181
INCONSISTENCY_CLOSED	198	198
REVIEW_RESOLVED	14	14
SUBJECTS_MERGED	4	4

The literal and URI counts were measured over every ACTIVE claim, so they are reproduced by the third command, which lifts the default exclusions of document-structure and non-asserted claims. The report also counts three SUBJECT_ALIASED events the first measurement did not look for, so its decision total as of 2026-10-01 is 400 against the 397 the four types sum to.

Today's store

Header figure	2026-10-02
Subjects	9,850
Aligned to any namespace	2,945
Classed	2,363, all in artifact:
ACTIVE structured claims	13,817
Claims in view	13,387
Literal / URI / token objects	13,243 / 144 / 0
Text / numeric / dated values	12,452 / 671 / 120
Distinct predicates	7,504
Canonical predicates	6,758
Canonical predicates with one claim	5,381
Share of claims under the 20 commonest canonical predicates	15.1%
Modelling decisions	402

Alignment by namespace and link band:

Namespace	Subjects	Asserted	Scored	Unscored	Suppressed
artifact	2,363	2,363	0	0	0
wikidata	582	3	340	41	198

The 6,758 canonical forms differ by one from the 6,759 the canonicalisation page reports for the core stem on the same listing. That page measured the stem before run was treated as its own participle; the report uses the shipped function.

The co-occurrence signal

The subject class a predicate attaches to is the evidence for its domain. The commonest canonical predicates on each class of subject, by claims:

Class	Claims	Commonest predicates
artifact:record	1,910	has status (138), covers (53), concerns (51), is (40), addresses (31)
artifact:file	1,224	contains (60), is located at (39), implements (37), is (25), requires (15)
artifact:symbol	379	has status (19), reports (11), returns (8), is defined in (7)
artifact:command	212	contains count (12), uses (5), supports flag (4), supports option (3)
Unclassed subject	7,409	is (205), passed (97), contains (97), is located at (71)
No resolved subject	2,253	is (81), requires (45), covers (25), returns (24)

Reading

• The evidence is real today and needs no new machinery. Records carry status and scope, files carry containment and location, commands carry flags. A reviewer could write each of those down as a domain from this table alone.
• Most claims attach to no class. 9,662 of the 13,387 claims in view are about an unclassed subject or none, so the co-occurrence signal covers about 28% of the store. Wikidata-resolved subjects record no class, which is the largest single gap.
• Alignment is thin and mostly estimated. 579 of the 582 Wikidata links carry a confidence below 1, and 198 of them are below the suppression threshold. An export that wrote them all as identity would assert what the resolver only estimated.
• Values are free text. 93% of values are text, and no literal in the store carries a datatype: the 671 numeric and 120 dated values are untyped literals whose form says what a datatype would.